PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning Based Automatic Modulation Recognition: Models, Datasets, and Challenges

Fuxin Zhang, Chunbo Luo, Jialang Xu, Yang Luo, FuChun Zheng

arXiv:2207.09647v1eess.SP

TL;DR

AMR must identify modulation schemes without prior information, but DL-AMR deployment raises concerns about complexity, explainability, and data quality. The paper reviews models and datasets, compares representative SISO approaches, and examines MIMO with precoding. Across four benchmark datasets, the experiments report high recognition accuracies, while the review highlights limited open benchmarks and the need for real-world data.

  • Problem

    AMR requires modulation identification without prior information, while DL-AMR practical deployment is constrained by concerns about complexity, explainability, and benchmark data.

  • Method

    The paper reviews DL-AMR models and datasets, experimentally compares representative SISO models under unified settings, and evaluates applying AMR to MIMO systems with precoding.

  • Results

    The highest reported accuracies across four datasets were 92.05% on RML2016.10a, 94% on RML2016.10b, 98.39% on RML2018.01a, and 100% on HisarMod2019.1.

  • Takeaways & Limitations

    The review provides model and dataset comparisons for practitioners and identifies unresolved DL-AMR challenges and future research opportunities.

  • Takeaways & Limitations

    Only a few open-source AMR datasets exist, and some simulate channels rather than collecting signals from real wireless environments.

Abstract

from arXiv · show

Automatic modulation recognition (AMR) detects the modulation scheme of the received signals for further signal processing without needing prior information, and provides the essential function when such information is missing. Recent breakthroughs in deep learning (DL) have laid the foundation for developing high-performance DL-AMR approaches for communications systems. Comparing with traditional modulation detection methods, DL-AMR approaches have achieved promising performance including high recognition accuracy and low false alarms due to the strong feature extraction and classification abilities of deep neural networks. Despite the promising potential, DL-AMR approaches also bring concerns to complexity and explainability, which affect the practical deployment in wireless communications systems. This paper aims to present a review of the current DL-AMR research, with a focus on appropriate DL models and benchmark datasets. We further provide comprehensive experiments to compare the state of the art models for single-input-single-output (SISO) systems from both accuracy and complexity perspectives, and propose to apply DL-AMR in the new multiple-input-multiple-output (MIMO) scenario with precoding. Finally, existing challenges and possible future research directions are discussed.

1. Introduction

AMR identifies wireless-signal modulation without prior information, supporting several radio applications while coping with channel and hardware distortions. This paper reviews DL-AMR models and datasets, compares representative methods experimentally, examines MIMO with precoding, and discusses challenges and future directions.

  • Motivation: AMR detects modulation schemes automatically without prior information and supports cognitive radio, spectrum sensing, surveillance, and interference identification.It operates as an essential step between signal detection and demodulation.
  • Motivation: Noise, multipath fading, shadow fading, frequency and sample-rate offsets, and hardware effects can distort signals and hinder modulation discrimination.
  • Deep learning motivation: DL-AMR research is motivated by deep networks’ feature-extraction and classification capabilities, with pioneering methods often outperforming traditional LB-AMR and FB-AMR approaches.
  • Paper scope: The paper reviews DL-AMR models and benchmark datasets, compares state-of-the-art SISO models under unified settings, and proposes AMR for MIMO communications with precoding.
  • Paper scope: The review also identifies remaining AMR challenges and discusses future research opportunities.

2. DL models for AMR in SISO systems

DL-AMR in SISO systems typically preprocesses received signals, extracts features, and classifies modulation schemes using diverse neural-network architectures. The reviewed models span unsupervised, DNN, CNN, RNN, and hybrid approaches, with datasets differing in scale, modulation content, and channel characteristics.

  • Signal model: The received signal is modeled after channel transmission and sampling, with channel gain, frequency and phase offsets, AWGN, and I/Q representation used for recognition.The task is to identify the transmitted modulation scheme from the received samples despite structural distortion.
  • General structure: DL-AMR generally comprises preprocessing, feature extraction, and modulation classification, implemented end-to-end or with traditional feature extraction preceding model training.
  • Unsupervised models: Unsupervised AMR includes DBN and autoencoder approaches; DBNs use stacked restricted Boltzmann machines and can combine spectral correlation functions with learned patterns.
  • DNN models: DNN-based AMR uses fully connected feedforward networks to classify modulation schemes from statistical features such as high-order cumulants, kurtosis, and skewness.
  • CNN-based models: CNN-based AMR uses raw I/Q data or preprocessed representations, including high-order cumulants, spectrum images, and constellation diagrams.Constellation-based inputs support CNN processing of signal representations, including filtered constellation-density matrices.
  • CNN-based models: Efficient CNN design reduces filters and trainable parameters to target low latency, with one reported model achieving processing time below 0.01 ms for B5G requirements.
  • RNN and hybrid models: RNN models learn temporal correlations, while hybrid CNN-RNN architectures combine spatial and temporal processing to address the limits of using either type alone.Examples include GRU-based RNNs and CLDNN models combining CNN and LSTM layers.
  • Datasets: The main SISO AMR datasets vary in size, SNR range, modulation coverage, and channel characteristics, with RML2016.10a commonly used as a benchmark.The RML datasets simulate effects including frequency and sample-rate offsets, AWGN, multipath, and fading.

3. AMR open source datasets for SISO systems

AMR datasets support model training, testing, and evaluation, but differ in scale, modulation coverage, signal length, and realism. Simulated datasets are efficient yet may not capture the complexity of real-world channels.

  • AMR datasets underpin deep-learning model training, testing, and evaluation.
  • RML2016.10a is a widely used benchmark generated with GNU Radio and designed to include common time-varying channel effects.
  • RML2016.10b provides more data, while RML2018.01a expands to 24 modulation schemes and lengthens samples from 128 to 1024, increasing computing demands.
  • HisarMod2019.1 contains 26 modulation types, but its better transmission environment makes the schemes easier to identify.
  • Simulated datasets can generate many samples quickly, but real-channel complexity and dynamics motivate training or fine-tuning with real-world data.

4. DL models for AMR in MIMO systems

MIMO DL-AMR exploits multiple received antenna signals and has been extended to systems with channel equalization, coding, transfer learning, and precoding. The adopted precoding setup uses channel SVD and receiver-side transformation before modulation recognition.

  • 4.2. The adopted MIMO system with precoding: The received MIMO signal is modeled as the channel-transformed transmitted vector plus additive white Gaussian noise.
  • 4.1. Introduction of the MIMO DL-AMR systems: MIMO systems jointly use signals from multiple antennas to determine the modulation scheme.
  • 4.1. Introduction of the MIMO DL-AMR systems: Existing MIMO DL-AMR methods include per-antenna CNN voting, CNN-based zero-forcing equalization, STBC-aware SAE-DNN models, and transfer-learning semi-supervised models.
  • 4.2. The adopted MIMO system with precoding: The paper experimentally tests DL-AMR in a MIMO communications system containing precoding.
  • 4.2. The adopted MIMO system with precoding: The adopted system assumes a flat-fading channel with N_t transmitter antennas and N_r receiver antennas, with channel knowledge available at both sides.
  • 4.2. The adopted MIMO system with precoding: SVD precoding decomposes the channel into unitary matrices and ordered singular values, providing the basis for transmitter and receiver processing.

5. Experimental comparison for SISO system

The SISO experiments compare state-of-the-art DL-AMR models across four benchmark datasets using common settings, examining accuracy, complexity, convergence, and confusion patterns. Results show dataset-dependent strengths, with recurrent models favored on RML datasets and simpler convolutional models effective on HisarMod2019.1.

  • Recognition accuracy: 92.05% on RML2016.10a, 94% on RML2016.10b, 98.39% on RML2018.01a, and 100% on HisarMod2019.1 were the highest reported accuracies under the shared settings.The maxima were achieved by MCLDNN at 10dB, LSTM at 18dB, LSTM at 22dB, and CNN1 and DenseNet at high SNR, respectively.
  • Recognition accuracy: RNN-based models including LSTM, GRU, PET-CGDNN, and MCLDNN show clear accuracy advantages on RML datasets, where temporal features support recognition.The passage attributes this pattern to temporal information in RML samples and the ability of RNN layers to extract it.
  • Recognition accuracy: CNN1 performs better on HisarMod2019.1, while deeper or more complex models may experience local optimality and degradation on that relatively ideal dataset.CNN2 can match or exceed CNN1 on certain RML tasks with fewer parameters, whereas ResNet and DenseNet perform well on HisarMod2019.1 but not equivalently on RML datasets.
  • Complexity and convergence: All models converge quickly and stabilize beyond 40 epochs, with MCLDNN, PET-CGDNN, LSTM, and GRU showing the best convergence trends across three RML datasets.The experiments use validation loss to assess convergence, while the study keeps hyperparameters fixed for fair comparison despite the larger RML2018.01a dataset.

6. Experimental comparison for the adopted MIMO system

The MIMO experiments test whether representative SISO DL-AMR models remain applicable when precoding converts multiple received streams into a single data stream. CLDNN achieves the highest overall accuracy among the tested models, and more antennas improve recognition accuracy.

  • Experimental design: The MIMO study compares representative CNN, LSTM, GRU, and CLDNN models to verify DL-AMR feasibility with precoding rather than introduce a new recognition architecture.Experiments use 4 × 2, 16 × 4, and 64 × 16 transmit–receive antenna configurations.
  • Recognition accuracy: CLDNN has the highest overall recognition accuracy, while CNN has the lowest overall recognition accuracy among the four tested models.The comparison is reported across the evaluated antenna configurations in Fig. 7.
  • Recognition accuracy: Increasing the number of antennas improves recognition accuracy for each model in the tested precoded MIMO configurations.The configurations are N_t=4, N_r=2; N_t=16, N_r=4; and N_t=64, N_r=16.
  • Applicability to MIMO: SISO DL-AMR methods can be applied to the tested precoded MIMO system because precoding produces one complex data stream requiring similar network structures.Other factors, including channel-estimation error and precoding schemes, remain identified areas for further investigation.

7. Challenges and future directions

DL-AMR still faces challenges in model design, data availability, generalization, efficiency, and explainability. Future directions include richer model architectures, standardized datasets, multi-channel processing, expert-knowledge fusion, and visualization.

  • Designing novel DL-AMR models: CNNs and RNNs have driven AMR breakthroughs, while GANs, attention mechanisms, transformers, and joint signal-processing/model design remain promising directions.GANs may expand training data, attention may focus on relevant features, and transformers may exploit the sequential nature of AMR signals.
  • Model optimization: Deeper models can improve accuracy but may impose excessive delays and exceed the memory, computing, and energy limits of IoT devices.Pruning, knowledge distillation, and weight quantization are proposed to reduce complexity while maintaining performance.
  • Model generalisation: Most DL-AMR models are trained under specific channel conditions, motivating transfer learning, zero-shot and few-shot learning, and defenses against adversarial attacks.Transfer learning can fine-tune models trained on offline datasets with real-world online data.
  • Data and datasets: Only a few open-source AMR datasets exist, and the field lacks solid common benchmarks for comparing and improving models.The paper identifies a need for higher-quality datasets, unified benchmarking, data cleaning and augmentation, and standard MIMO datasets.
  • Benchmark AMR models, datasets, and data challenges: Multi-channel inputs and feature fusion have achieved higher AMR accuracy than sophisticated single-input models, but increase computational complexity and may impair convergence.Parallel processing on GPUs could reduce computational costs, while new loss functions or data-dependent weight adjustments may address convergence.
  • Data and datasets: Supervised AMR models require substantial labeled data, creating interest in semi-supervised learning because labeled datasets demand significant time and resources.Existing unsupervised approaches such as DBNs and autoencoders have not achieved comparatively distinctive results.
  • Designing novel DL-AMR models: Combining DL with wireless-communications expertise, including cumulants, handcrafted features, spectra, constellation diagrams, filtering, and denoising, could improve AMR.
  • Model optimization and visualisation: Low-SNR inputs make DL-AMR features difficult to assess visually, limiting model design and motivating visualization methods for optimization and explainability.

8. Conclusion

The paper reviews DL-AMR models and datasets, identifies unresolved problems, and proposes possible solutions supported by extensive experimental comparisons. It positions these insights as relevant to future cognitive communications and distributed networked devices in B5G/6G networks.

  • The paper synthesizes DL-AMR models, benchmark datasets, unresolved problems, and possible solutions through extensive experimental comparisons.
  • The reviewed challenges and proposed directions are framed for cognitive communications systems and distributed networked devices in future B5G/6G networks.
Loading 2207.09647v1…